30.7 Conclusions

This chapter presents a new application of fully abundance-constrained LSMA to subpixel target size estimation. The idea is to apply an unsupervised target detection algorithm, ATGP, to find subpixel targets of interest, and then implement FCLS to estimate the abundance fractions of subpixel targets present in the image, and finally use the obtained abundance fractions to calculate their sizes. For such an approach to be effective, an accurate estimate of abundance fraction for a subpixel target is required. In this case, a fully constrained abundance LSMA such as FCLS is implemented for this purpose. Despite that abundance-constrained linear unmixing has been studied extensively for material quantification, the issue of subpixel target size estimation investigated in this chapter has never been explored in the past. Four LSMA methods are used for validation, which are an unconstrained method, LSOSP, two partially constrained least-squares methods, SCLS, NCLS, and a fully constrained method, FCLS. As demonstrated in simulated and real image experiments, the need of the fully abundance-constrained methods is evident when the target size is smaller than GSD. The target estimation error is increased as the target size is decreased. In addition to subpixel target size estimation this chapter also explores another application, the problem of concealed target detection, and further develops a computer automated method for detecting and classifying unknown concealed targets, to be called computer-aided detection and classification algorithm (CADCA). Since the targets of interest are either shaded by natural background or hidden underneath man-made objects, a band ratio transformation is used to reduce the effects caused by topographic aspects or differential illumination. In order to select appropriate band images to be used for band ratio, a BBOPC approach is also proposed for this purpose. Since the unknown concealed targets is now uncovered by the custom-designed band ratio transformation, an unsupervised target detection algorithm, ATGP is readily applied. It is worth noting that the criterion for ATGP is based on orthogonal subspace projection, which is also used for BBOPC. Such consistent design principle allows CADCA to achieve its best possible overall performance.